AREX: Towards a Recursively Self-Improving Agent for Deep Research

Researchers introduced AREX, a deep research agent that alternates evidence gathering with constraint-based self-auditing.

It uses a compact internal state to track unresolved claims and recursively improve long-horizon answers.

AREX was trained on synthetic verified tasks with reinforcement learning and agentic mid-training.

The system outperformed comparable baselines on several research and reasoning benchmarks, including BrowseComp, WideSearch, DeepSearchQA, and Humanity's Last Exam.

The work points to more reliable self-improving agents for complex research tasks where verification is easier than discovery.